A study of the influence of REV variability in double scale FEM×DEM analysis

Abstract

In this work, the consequences of using several different Discrete Element granular assemblies for the representation of the microscale structure, in the framework of multiscale modelling, have been investigated. The adopted modelling approach couples, through computational homogenization, a macroscale continuum with microscale discrete simulations. Several granular assemblies were used depending on the location in the macroscale Finite Element mesh. The different assemblies were prepared independently as being representative of the same material but their geometrical differences imply slight differences in their response to mechanical loading. The role played by the micro-assemblies, with weaker macroscopic mechanical properties, on the initiation of strain localization in biaxial compression tests is demonstrated and illustrated by numerical modelling of different macroscale configurations. This article is protected by copyright. All rights reserved.

Full text

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING
Int. J. Numer. Meth. Engng (2016)
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/nme.5202
A study of the influence of REV variability in double-scale
FEM DEM analysis
Ghassan Shahin1,2, Jacques Desrues2,1,*,† , Stefano Dal Pont1,2, Gaël Combe1,2
and Albert Argilaga1,2
1Université Grenoble Alpes, 3SR, F-38000 Grenoble, France
2CNRS, 3SR, F-38000 Grenoble, France
SUMMARY
In this work, the consequences of using several different discrete element granular assemblies for the rep-
resentation of the microscale structure, in the framework of multiscale modeling, have been investigated.
The adopted modeling approach couples, through computational homogenization, a macroscale continuum
with microscale discrete simulations. Several granular assemblies were used depending on the location in
the macroscale finite element mesh. The different assemblies were prepared independently as being repre-
sentative of the same material, but their geometrical differences imply slight differences in their response to
mechanical loading. The role played by the micro-assemblies, with weaker macroscopic mechanical prop-
erties, on the initiation of strain localization in biaxial compression tests is demonstrated and illustrated by
numerical modeling of different macroscale configurations. Copyright © 2016 John Wiley & Sons, Ltd.
Received 5 August 2015; Revised 15 December 2015; Accepted 23 December 2015
KEY WORDS: multiscale modeling; FEM DEM; representative elementary volume (REV); damage
analysis; strain localization; granular material
1. INTRODUCTION
Multiscale modeling of granular materials has received an increasing attention in the recent years
[1–6] as these approaches allow to naturally embed a refined description of the complexity of the
material into a full structural engineering problem. Moreover, these techniques allow to efficiently
overcome the common drawbacks of conventional modeling strategies [7, 8]. In this regard, finite
elements-based approaches allow to treat large-scale problems; however, the application of the
method to complex materials (such as geomaterials) may be unsatisfactory because of the difficulties
arising in the definition of an adequate mechanical constitutive law. On the other hand, the discrete
element method (DEM) [9–11] is usually considered particularly effective to capture the nature of
geomaterials and properly describe their constitutive behavior. However, the need for a large num-
ber of particles to simulate a full-engineering problem imposes some restrictions to the application
of this method [12].
Recently, the two methods have been coupled in the framework of a multiscale numerical homog-
enization approach to exploit the efficiency of FEM at solving boundary value problems at structural
level and the capability of DEM to capture complex material behaviors. Such an approach is usually
known as FEM DEM method. Several multiscale modeling schemes have been proposed. Kaneko
et al. [1, 13] employed the mathematical homogenization theory for the construction of a multiscale
modeling approach. Miehe et al. [3, 14] proposed an original scheme for quasi-static homog-
enization of granular microstructures and its embedding into a two-scale modeling framework.
*Correspondence to: Jacques Desrues, Laboratoire 3SR, Domaine Universitaire, BP53, 38041 Grenoble Cedex 9, France.
†E-mail: jacques.desrues@3sr-grenoble.fr
Copyright © 2016 John Wiley & Sons, Ltd.

G. SHAHIN ET AL.
Meier et al. [2, 15] built an FEM DEM scheme employing the continuum-based Taylor assump-
tion at the microscale level. In these methods, the DEM is used to derive an equivalent mechanical
law for the microstructure behavior. Andrade et al. [5] and Avial et al. [16] developed a discrete-
continuum approach based on a numerical homogenization scheme in which friction and dilatation
at the microstructure are transferred to the elasto-plasticity continuum at the macroscale level.
Nitka et al. [4, 17], Nguyen et al. [18, 19], Guo and Zhao [6, 20], and Desrues et al. [21] applied
the concept of the representative elementary volume (REV) to build a fully coupled multiscale
FEM DEM approach based on computational homogenization. The stress state at the macroscale
level is obtained, for each Gauss point, from the associated DEM granular assembly (REV).
The efficiency of the method has been explored through several study cases, basically consist-
ing of monotonic compression biaxial test [1, 3, 5, 16, 19, 20, 22], cyclic simple shear test [20],
hollow tube test [21], and slope stability test [15], in addition to the retaining wall problem and
the classical footing problem [23]. These numerical experiments were run using both cohesive and
disperse materials and considering only a single DEM granular assembly to represent the material
microstructure constitutive behavior over the whole specimen. However, computerized tomography
scans (e.g., Figure 1) reveal a strong randomness in the distribution of the particles over a sample
volume, thus questioning the assumption of a single REV. According to experimental observation,
the use of a single microstructure volume periodically distributed over the entire specimen may lead
to the loss of some distinct features of the considered geomaterial.
The scope of the paper is to move towards a more realistic representation of geomaterials by
considering, in the framework of a FEMDEM approach, different REVs at the microstructural
level. The used REVs are representative of the same material, (i.e., the same material properties with
random realization of the same grain size distribution), but they differ with respect to their geometry
(i.e., the positioning of the particles is different), as suggested by the experimental evidence.
The paper is structured as follows. Section 2 briefly summarizes the adopted multiscale
FEM DEM modeling approach. Section 3 presents the preparation algorithm of the granular
assemblies and provides identification and selection criteria of the granular assemblies. Section 4
describes the FEM DEM modeling details and the biaxial test procedures. Section 5 shows pure
DEM and homogeneous FEM DEM simulations. Section 6 proceeds towards the focal issue in this
work to carry out a set of inhomogeneous FEM DEM simulations. Section 7 uses the DEM REV
of the weak mechanical properties to introduce material imperfection into homogeneous simulation.
The work is then concluded, and some perspectives are presented.
Figure 1. A computed tomography scan of dry sand laboratory sample [24]. Although it was prepared using
the same material, this homogeneous sample implies strong fluctuations in microstructures geometry.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
2. MULTISCALE FEM DEM MODELING APPROACH
In this work, the multiscale FEM DEM modeling approach proposed by Nitka et al. [4] and then
improved in Nguyen et al. [18, 19] and Desrues et al. [21] has been adopted as a general frame-
work. In this method, a quasi-static finite strain macroscopic continuum is coupled with discrete
microstructures defining the constitutive behavior of the material. The macrostructure boundary
value problem is described by means of FEM, whereas the local constitutive relation at the Gauss
point level is determined by a DEM granular assembly that is considered as an REV. The local
REV is representative of the material behavior, and each REV has its own stress–strain evolution
(like a standard Gauss point in FEM). In the sequel, the modeling approach is restricted to the 2D
simulation at both macroscopic and microscopic levels.
The constitutive law at the macrostructural level gives the stress state ij .t/ as a function of the
displacement gradient history hkl:
ij .t/ Dt¹hkl . /;  2.0; t /º(1)
The DEM–FEM coupling is obtained through a computational homogenization method follow-
ing the typical step shown in Figure 2. As in standard finite elements, the displacement gradient
increment ıhkl is applied at the Gauss point level to obtain the corresponding stress state. In the
FEM DEM scheme, the displacement gradient increment acts as an updated boundary condition
on the discrete element REV at the Gauss point level. The subsequent discrete element numerical
simulation acts as a material constitutive relation by returning the new stress state at the FE-macro
level. The updated Cauchy stress state is obtained through the standard Weber [25] homogenization
formulation:
ij D1
S:X
.p;q/2c
fq=p
i˝lp=q
j(2)
where Sis the area of the assembly, fq=p and lp=q are, respectively, the interparticle forces and the
branch links of the p; q grains. cis the set containing all the contacts in the granular assembly.
Computational homogenization techniques typically rely on a scales-separation assumption
[8, 26, 27], thus requiring that the characteristic lengths of the spatial variation of macroscopic fields
are much larger than the characteristic size of the embedded microstructure. This assumption holds
as long as localization does not occur. As soon as a localization band develops, the microstructural
element is no longer statistically representative (strictly speaking, a generalized Microstructural
Element Volume rather than an RVE should be introduced). The interested reader should refer to,
for example, Reference [26] or, more recently, Reference [27].
Figure 2. Computational homogenisation scheme [18]. FEM, Finite element method.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Table I. The equilibrium criteria and the six levels of equilibrium conditions of DEM computation.
Level of equilibrium 0 1 2345
Min number of iterations 10000 10000 10000 10000 10000 10000
Max kinetic energy 1.0E-2 1.0E-5 1.0E-8 1.0E-8 1.0E-8 1.0E-8
Max proportion of sliding contacts 10% 10% 10% 10% 1% 0%
Max proportion of grains with single contact 10% 10% 10% 10% 1% 0%
Max resultant/minimal forces acting on each grain 1.0E-2 1.0E-2 1.0E-2 1.0E-2 1.0E-2 1.0E-6
The used FEM code [28] is a finite strain implementation of the standard finite element
method [29], and the DEM code is an in-house implementation of the standard discrete element
method [9]. The DEM model assumes that all grains interact under linear-elastic normal and tan-
gential stiffness. Coulomb-type friction is introduced into contact system, and energy is dissipated
through a viscous damping mechanism.
2.1. FEM algorithm
The FEM solution is integrated by an adaptive step-by-step strategy in which the solution of a typical
step is obtained by means of Newton–Raphson method [30]. The Newton–Raphson method requires
the determination of the consistent tangent operator (CTO) Cij k l for each step in order to integrate
the solution.
Cij k l Ddij
dhkl
(3)
The CTO is numerically computed for the first quasi-elastic steps using the perturbation
method [19]. When the material behavior becomes strongly nonlinear (at about 0.5% axial strain),
the CTO computation becomes numerically less efficient. Therefore, a different strategy is adopted,
and the tangent operator is considered as the average of the tangent operators of the first 10 steps.
As an auxiliary operator is used, the quadratic convergence, which is supposed to be obtained
with a Newton–Raphson method, is lost, and a larger number of iterations will be required.
Such a modified Newton–Raphson approach, however, proved to be more robust than the full
Newton’s method [18] based on the computations of the CTO. This iterative scheme is associated
with two different convergence criteria based on out-of-balance forces and displacement decrement
(Appendix A).
2.2. DEM algorithm
The displacement gradient increment ıhkl is transferred from the FE macroscale and applied to
the DE microstructure through an explicit step-by-step integration scheme. The application of ıhkl
is conducted with fixed time step ıt that is computed based on the dynamic properties of the
contacted particles. After the application of the strain, the assembly is submitted to a relaxation
phase that ends once the assembly reaches an equilibrium state. The equilibrium of the assem-
bly is estimated through the five criteria that are illustrated in Table I. At the granular level, the
solution of motion equation is obtained by means of a predictor–corrector [31] integration method
of order 3.
3. DEM REVS SIMULATIONS
In the framework of the FEMDEM modeling approach, the DEM granular assembly is used to
define the constitutive response of the considered material. Preparation procedures of the DEM
assembly, to be used as REV in this process, are of great importance because of their strong effects
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
on the mechanical properties of the final assembly [10, 11]. Hereafter, a detailed description of the
followed preparation procedures is presented, and a set of granular assemblies is generated accord-
ingly. A subset of granular assemblies is then selected, and a biaxial test simulation is proposed, in
Section 5.1, to explore their mechanical behavior.
3.1. The DEM model
A standard DEM model [10, 11] has been used in this work. The model assumes that all rigid circular
grains interact when they are in contact. The normal contact force is given as fnDKnı,where
Knis the normal contact stiffness and ıis overlapping magnitude (ı<0when the contact occurs).
The tangential force increment is given as ıf tDKt:ıUt,whereKtis the tangential stiffness and
ıUtis an increment of tangential relative displacement computed at the contact point. The total
tangential force, ft, is the sums over the time step ıt. The Coulomb friction coefficient is used to
limit ftsuch that jftj6f n,whereis the contact angle of friction. The kinetic energy of the
system is dissipated through viscous dampers introduced to each contact system. If miand m
jare
the mass of the two particles in contact, damping coefficient is defined as a proportion of the critical
damping, , which is computed using the relation D2Knmimj
miCmj0:5.
3.2. DEM REV generation
The assembly is formed of 400 particles, consistently with the guidelines provided by Nguyen
et al. [18, 19] and Guo and Zhao [6, 20]. This choice constitutes a good compromise between
the computational cost and the problem stability. In order to obtain homogeneous assemblies, the
preparation procedures were performed in the absence of gravity, and periodic boundary condi-
tions were retained in the horizontal and vertical directions [10, 11]. In this work, the simulation is
focused on a dense frictional granular material (without cohesion). To obtain a dense assembly, the
inter-particle friction coefficient has been set to zero during the preparation step as the particles can
slide and fill the volume as much as possible. In addition, the grain size range has been taken as
Rmax =Rmin D2:5.
Discrete element method granular assemblies have been generated following the algorithm pro-
posed in [10] and [11]. Starting from a given particle size range and a given number of particles,
a random generator places the particles on a regular grid of unit length modulus, (Figure 3(a)).
The particles are then submitted to a random velocity field. The particles move and interact as
rigid bodies within a fixed size container. No energy dissipation is introduced during this process.
When particles have been shaken enough (i.e., each particle is displaced cumulatively 100 times its
Figure 3. The three stages of the granular assembly preparation with periodic boundary conditions. (a) The
ordered configuration of the assembly; (b) the granular assembly after shaking; and (c) the final configuration
after the isotropic compaction of the granular assembly.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Table II. The given values of DEM mechanical parameters.
Parameter Value
Normal stiffness coefficient DKn=P 1000
Tangent/normal stiffness coefficients Kt=Kn1
Coulomb friction 0.5
Viscous damping coefficient 0.95
is the normal stiffness coefficient normalized by confinement
pressure P. Damping coefficient is a proportion of, , the critical
damping of contact system.
diameter), their velocity is set to zero, (Figure 3(b)). The resultant granular packing, which has a
gas-like configuration, is then subjected to a strain-controlled isotropic compression phase leading
to the desired granular assembly, (Figure 3(c)).
Following these procedures, more than 20 granular assemblies have been generated using the
material mechanical properties presented in Table II. The proposed procedure allows to generate
REVs characterized by major differences in particles positioning but only minor fluctuations in par-
ticle size distribution. The generated REVs can therefore be considered as representative of the same
material. Because the objective of this study is to investigate the consequences of using different
DEM REVs, each of these granular assemblies will be given an identity based on the criteria that
are presented in the following section.
3.3. Granular assembly identity and selection criteria
Radjai and Dubois [10] introduced a number of criteria to determine the identity of a given granu-
lar assembly. They defined a set of internal variables that distinguish one granular assembly from
another. Among these variables, packing fraction (PF) and coordination number (CN) have been
retained hereafter. The PF describes the proportion of solids volume to the overall assembly vol-
ume, whereas the CN represents the average number of inter-particles contacts per particle in the
packing. If Ncis the total number of contacts, and nis the number of particles in contact, then
CN D2Nc=n.
The generated assemblies exhibit a Gaussian distribution of PF and CN that can be formulated as
PF D0:8157 ˙0:004 and CN D4:153 ˙0:015. The small deviation of PF and CN proves the
efficiency of the proposed procedure in generating granular assemblies representative of the same
material. Such a small deviation, however, excludes the typical correlation with the strength peak,
as the relation between strength peak and these two variables exhibits a random scattering within a
small window of variation. This scattering can significantly be affected by the number of grains; a
smaller number of grains promotes the role of geometry, as any small variation in the fabric would
result in significant changes in the strength peak. Therefore, the introduced randomness in geometry
factors at preparation process will show, in the next section, a clear influence on the strength peak
of the granular assemblies.
As aforementioned, the idea of using different granular assemblies is to represent the random geo-
metrical variations of the microstructure in soil sample, apart from any deviation in the mechanical
properties. Thus, the selection of the granular assemblies has to be conducted independently from
their mechanical properties. Given that the mechanical properties (strength peak) of the assemblies
exhibit a random scattering within a small variation of the internal variables, considering PF and CN
as selection criteria can offer such sort of independent process. The assemblies corresponding to the
mean value ˙a standard deviation of PF and CN have been selected. The assemblies corresponding
to the mean value of PF plus/minus a standard deviation are denoted by PF-A/PF-B, respectively.
Whereas, the assemblies corresponding to the mean value of CN plus/minus a standard deviation
are denoted CN-A/CN-B, respectively.
Another significant variable has to be considered. In the procedures, compaction process is con-
ducted as the normal strain components are imposed, but not the shear strain component. Instead,
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
zero shear strain is imposed (i.e., orthogonal wall driving). Even though, at the end of the prepa-
ration stage, the obtained assemblies always show a slight deviation from isotropy. Equation (4)
shows a comparison between the target stress state (left matrix) and the homogenized internal stress
state using Equation (2) (right matrix) for an arbitrarily chosen packing from the generated assem-
blies. This deviation from isotropy can be attributed to the finite number of grains in the granular
assemblies. Despite its small magnitude, this deviatoric stress component will affect the mechanical
behavior once introduced in the double-scale computations of FEM DEM scheme, as it introduces
an initial bias to simulation specimen [20]. In the following, this deviatoric stress component will
be termed as remaining shear stress (RS). To investigate the effects of RS on the interplay between
the microstructure and macrostructure, it has been considered as a third selection criterion. The
assemblies corresponding to both the maximum positive and the minimum positive RS have been
selected and denoted by RS-A and RS-B, respectively. In fact, the assembly associated to the mini-
mum RS has already been considered for the CN criterion. Replicating the use of the same assembly
might render some difficulties at reading the plots. Therefore, another one has been chosen, namely,
an assembly with RS magnitude close to zero. Table III shows the chosen assemblies and their
notation, in addition to the associated values of PF, CN, and RS.
10
01
¤1C0:07591
C0:07591 1 (4)
Table III. The selected DEM REVs based on the three criteria, PF, CN, and RS, for
microstructure representation in multiscale simulation.
PF, packing fraction; CN, coordination number; RS, remaining shear stress.
Figure 4. The finite element model and its boundary conditions used at the macroscale level of the multiscale
simulation specimens of aspect ratio 2 discretized using (510) 50 Q8 finite elements.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Table IV. The adopted computational parameters
values based on the performed sensitivity study.
IN LE PM EDjEF
1.0E-4 2 15.0E-6 0.01jNA
IN, inertial number; LE, level of equilibrium; PM,
perturbation magnitude; ED, displacement-based
convergence criterion; EF, force-based convergence
criterion.
4. FEM DEM MODELING SETUP
The biaxial test is the equivalent for 2D modeling of the so-called triaxial test widely used for the
characterization of granular media behavior [32, 33]. Herein, the idea is to use the FEM DEM
approach to develop a biaxial test simulation. A detailed description of the FEMDEM model and
the applied biaxial test procedures are presented hereafter.
4.1. The numerical specimen
The macroscale of the specimen to be modeled is discretized into 5 10 finite elements mesh (aspect
ratio 2). The eight-nodes quadrilateral element with four integration Gauss points (Q8) is retained
in the study. The finite element mesh and its boundary conditions are illustrated in Figure 4. Each
Gauss point in this mesh is associated with a specific DEM assembly. The six DEM assemblies
already presented in Table III will be used for this purpose.
4.2. Numerical experiment setup
The FEM DEM numerical study is focused on a standard monotonic compressive biaxial test. In
its initial state, the specimen is subjected to an isotropic compression. Thereafter, the horizontal
confinement pressure is held constant and the deviatoric stress is imposed through strain-controlled
vertical compressive loading, for more details see, for example, Desrues and Viggiani [32].
The numerical procedures of both FEM and DEM, presented in Section 2, imply the choice of
a set of computational parameters, which play a major role on the efficiency of the analysis: the
convergence rate at both macroscale and microscale. These parameters are the inertial number (IN)
and the level of equilibrium (LE) conditions for the DEM scheme, the perturbation magnitude, and
the precision level (EFand ED) for the FEM scheme. For more details about these four parameters,
see Appendix A. These parameters were identified via a sensitivity analysis that is presented in
Appendix A and summarized in Table IV.
5. HOMOGENEOUS MULTISCALE SIMULATION
In this part, several FEM DEM biaxial test simulations are performed using the REVs introduced
in Table III. A unique DEM REV is used in each simulation, thus obtaining an initially homogeneous
specimen having the same mechanical properties over the entire domain. Before conducting the
FEM DEM simulations, the behavior of the used DEM REVs is investigated by means of pure
DEM simulation of biaxial test. These simulations describe the behavior of the material itself, and
they will be used for comparison with the multiscale FEM DEM simulations that exhibit, after the
localization of strain field, structural behavior rather than material behavior.
5.1. Material behavior – pure DEM simulations
The behavior of the material used in our numerical experiments is first explored through a pure
DEM simulation. Each DEM assembly is subjected to a biaxial loading path up to 8% axial strain.
Figure 5 shows the global response of the six assemblies plotted as a function of the deviatoric
stress qnormalized by the confinement pressure 00 versus the applied axial strain. The deviatoric
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
Figure 5. Pure DEM simulations of monotonic compression biaxial test up to 8% axial strain performed on
six different DEM assemblies of identical material properties but different with respect to geometry (different
particles positions and grain size distribution). (a) PF-A and PF-B; (b) CN-A and CN-B; and (c) RS-A
and RS-B.
stress is computed as qD12,where1and 2are the vertical normal stress and the lateral
normal stress, respectively. The obtained results show that all cases exhibit, after a pre-peak smooth
response, global strength loss, which is a characteristic of strain softening behavior. The strong
fluctuations observed in the post-peak part of the curves can be attributed to the relatively small
number of particles (400 particles) that leads to a series of sudden contact rearrangements along the
deforming process. Despite these differences in the strength peak, all assemblies tend to a similar
level of residual strength (plateau around q=00 D0:9).
Though the six assemblies have the same DEM mechanical parameters (Table II) and differ only
with the random geometric factors (described in Section 3.3), the pure DEM simulations show that
they have different strength peaks; the resultant normalized deviatoric strength peaks vary from
1.3 to 1.8. This divergence in the granular assemblies behavior comes only from their geometry
differences. Subsequently, these observations show that geometry differences have strong effects on
the mechanical properties of the granular assembly.
5.2. FEM DEM simulation
Each REV is then used for the initially homogeneous FEM DEM simulations. The numerical spec-
imen is subjected to a biaxial loading path up to 8% axial strain. Figure 6 shows the global response
of the six FEM DEM specimens compared with the pure DEM simulations. In this figure, all the
cases show that each FEM DEM specimen has the same mechanical response as the associated
DEM REV up to the strength peak. In contrast, the post-peak response of FEM DEM specimens
show a clear divergence from the DEM simulation. Indeed, beyond the peak, strain localisation starts
to develop, which explains the divergence from material behavior. Moreover, while the response of
the DEM assembly strongly fluctuates, the FEM DEM specimen exhibits a smoother response.
This can be attributed to the averaging effect of the integration by the FEM numerical process
of the inherently noisy responses of the different REVs involved in the structure, especially those
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Figure 6. A comparison between the global responses of the homogeneous FEM DEM simulations with
the pure DEM simulations of the corresponding DEM REV in biaxial test simulations up to 8% axial strain.
Figure 7. Global responses of six homogeneous specimens in a biaxial test up to 8% axial strain simula-
tions. The used DEM REVs, for microscale, are identical with respect to material properties but different
concerning the geometry.
associated to the Gauss points that lie in the shear band zone. At the end of the test, the FEM DEM
specimen and the DEM assembly tend to a similar residual strength plateau.
Figure 7 compares the six FEM DEM simulations. It can be observed that even with different
strength peaks, all the FEM DEM simulations tend to a similar residual strength level (around
q=0D0:9). This result is consistent with what is classically observed in DEM modeling: for large
strain, the mechanical behavior of granular material is mainly ruled by the particle shapes and the
inter-granular angle of friction [34].
6. INHOMOGENEOUS MULTISCALE MODELING
So far, a single DEM assembly has been used in each FEMDEM simulation. The use of a unique
DEM REV over the entire mesh implies a uniform representation of the microstructure and leads to
an initially homogeneous specimen. However, the experimental evidence shows that a geomaterial
sample presents a variety of microscopic structures (Figure 1). The differences of the microstructure
in heterogeneous geomaterials result in a variability of the local mechanical properties, which might
play a major role in the resulting global behavior and the formation of the shear band. This part of
the work is devoted to this particular issue, which is the motivation of the whole study.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
6.1. Inhomogeneous simulations setup
The FEM DEM approach naturally offers the possibility of taking into account the heteroge-
neous nature of geomaterials. The variation of the local mechanical properties, originating from the
microstructure, is then introduced into an FEM DEM simulation by using several DEM assem-
blies, representative of the same material but with geometrical differences (Table III). This approach
allows to enhance the original model by embedding heterogeneities, while the average material
properties are the same. To simplify this investigation, the heterogeneity has been introduced in
the following FEM DEM simulations by using two different DEM REVs. The way the two DEM
REVs interplay with the macroscale continuum is investigated hereafter.
Based on the selection criteria introduced in Section 3.3, the set of the six DEM REVs was split
into three groups: PF group (PF-A and PF-B), CN group (CN-A and CN-B), and RS group (RS-A
and RS-B). Each group is referred as PFG, CNG, and RSG, respectively. In Table V, the granular
assemblies in each group were distinguished into strong and weak REVs according to their strength
peak (Section 5.1). The granular assembly with the higher strength peak is referred as strong REV
(i.e., PF-A, CN-B, and RS-B), whereas the assembly with the lower strength peak is referred as
weak REV (i.e., PF-B, CN-A, and RS-A) (Figure 5).
Each pair of these REVs is then used in different FEM DEM simulation. The in-plane distribu-
tion of the two DEM REVs is performed following two different patterns: a checkerboard pattern
and a random distribution pattern (Figure 8). Consequently, the six inhomogeneous specimens have
been used to run a set of biaxial test simulations up to 8% axial strain. Each FEM DEM simulation
will be identified by the name of the associated REVs group, namely, PFG, CNG, and RSG.
6.2. Results and discussion
Figure 9 shows the results obtained from PFG simulations with the checkerboard pattern (C-PFG)
and the random distribution pattern (R-PFG). Figure 9(a) shows the global response of these inho-
mogeneous specimens. The global response of the corresponding homogeneous specimens, PF-A
and PF-B, are presented in the same figure. This comparison shows that the inhomogeneous speci-
mens yield to identical pre-peak responses. In both cases, the strength peak of the inhomogeneous
specimen has been determined by the strength capacity of the weak REV (PF-B). Furthermore, the
inhomogeneous specimens show a clear divergence in the strain softening behavior; however, they
tend to a similar residual strength. Figure 9(b) and (c) displays the cumulative deviatoric strain field
Table V. The six DEM assemblies split into three different subsets. The dark gray
and the light gray refer to the strong and the weak REVs, respectively.
Figure 8. The checkerboard pattern and the random distribution pattern of the inhomogeneous specimens.
The dark gray refers to the elements that were associated with the strong REV, whereas the light gray refers
to the elements that were associated with the weak REV.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Figure 9. (a) The global response of the inhomogeneous specimens in which PF-A and PF-B have been
attached in the dark gray and light gray elements, respectively, following the checkerboard pattern (blue) and
the random distribution pattern (red). The global response of the homogeneous specimens corresponding to
the used DEM REVs are presented as well. (b) and (c) The cumulative deviatoric strain field corresponding
to the specimen of checkerboard pattern and random distribution pattern inhomogeneity, respectively, at the
end of the test.
Figure 10. (a) The global response of the inhomogeneous specimen in which CN-A and CN-B have been
attached in the dark gray and light gray elements, respectively, following the checkerboard pattern (blue) and
the random distribution pattern (red). The global response of the homogeneous specimens corresponding to
the used DEM REVs are presented as well. (b) and (c) The cumulative deviatoric strain field corresponding
to the specimen of checkerboard pattern and random distribution pattern inhomogeneity, respectively, at the
end of the test.
Figure 11. (a) The global response of the inhomogeneous specimen in which RS-A and RS-B have been
attached in the dark gray and light gray elements, respectively, following the checkerboard pattern (blue) and
the random distribution pattern (red). The global response of the homogeneous specimens corresponding to
the used DEM REVs are presented as well. (b) and (c) The cumulative deviatoric strain field corresponding
to the specimen of checkerboard pattern and random distribution pattern inhomogeneity, respectively, at the
end of the test.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
at 8% axial strain (the end of the simulation) of the specimens with the checkerboard pattern and
the random distribution pattern, respectively. In Figure 9(b), it should be noted that the shear band,
although it may seem to reach the left boundary of the specimen at the upper-left corner, is in fact
essentially aligned with the line of light gray elements below this corner. As a result, both cases
demonstrate that the strain field has localized in the elements that were associated with the weak
REV (PF-B).
The results obtained from the CNG simulations with the checkerboard pattern (C-CNG) and
the random distribution pattern (R-CNG) are presented in Figure 10. These results confirm the
aforementioned observations regarding the identical pre-peak response and the dominant role of
the weak REV (CN-A) at defining the strength peak of the inhomogeneous specimen. The great
dispersion of the mechanical behavior of the REVs, used in this set of experiments (Figure 5),
provides more pronounced evidence of such role. Also, they corroborate the role of the weak REV
at the definition of the shear band pattern.
Figure 11 shows the results of the third set of experiments in which the RSG assemblies were
retained. The global response of the inhomogeneous specimens compared with the corresponding
homogeneous simulations are presented in Figure 11(a). This comparison confirms the observa-
tions made earlier concerning the capability of the weak REV (RS-A) at defining the strength peak
of the inhomogeneous specimen. Also, it confirms that, after a divergent strain softening behavior,
the inhomogeneous specimen shows a similar residual strength level. Figure 11(b) and (c) dis-
plays the cumulative deviatoric strain field at 8% axial strain of the inhomogeneous specimens of
checkerboard pattern and random distribution pattern, respectively. Both cases have demonstrated
similar observations regarding the shear band pattern that is determined by an alignment of elements
attached with the weak REV. Moreover, another feature of the shear band pattern has been obtained
in the simulation with checkerboard inhomogeneity: the strain field has localized in two different
shear bands reflecting the periodicity of inhomogeneity.
The homogeneous simulations of both CNG and RSG have shown clear differences in the strain
softening behavior; while the homogeneous CN-B and RS-B simulations exhibit brittle behav-
ior, the CN-A and RS-A simulations are rather ductile. However, the results obtained from the
CNG and RSG simulations set reveal that the inhomogeneous models tend to exhibit a more
ductile behavior. This matter can be attributed to the role of the weak REV, which is of more
ductile behavior, at triggering strain localization and then defining the shear band pattern in the
inhomogeneous simulation.
As a conclusion, the use of two different DEM REVs for the microstructure representation helps
the specimen to define its shear band pattern that develops at an alignment of elements associated
with the weak REV. Moreover, the weak REV triggers strain localization and consequently plays a
major role in the definition of the specimen strength peak.
7. MATERIAL IMPERFECTION SIMULATION
The capability of using several REVs to capture the intrinsic heterogeneity of geomaterials have
been explored in the previous section. In what follows, a material imperfection, represented by the
weak REV, is introduced into homogenous simulation.
Experimental evidence show that introducing material imperfection into soil specimen facilitates
triggering strain localization and then this imperfection becomes a crucial path of the consequent
shear band [32]. The efficiency of simulating this experiment will be investigated using the weak
REV to introduce material imperfection into homogeneous specimen. The granular assemblies set,
CNG, has been chosen considering that CN-A and CN-B have shown the greatest dispersion in the
mechanical properties (strength peak) (Figure 5(b)). The imperfection is introduced into the speci-
men by changing the DEM REV in a single element. The remaining 49 elements are associated with
the CN-B. Figure 12(b) depicts the distribution of the DEM REVs over the FE mesh. Hereafter, this
specimen is referred as IMP. As in the previous cases, the behavior of the IMP has been investigated
through a biaxial test simulation up to 8% axial strain. Figure 12(a) shows the global response of
the IMP specimen; the global responses of the corresponding homogeneous specimens, CN-A and
CN-B, have been presented in the same figure as well.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Figure 12. A comparison between homogeneous specimen and a specimen that have been introduced
material imperfection by attaching an REV of weaker strength peak at a single element: (a) compares the
global responses of the imperfect specimen (IMP) and the corresponding homogeneous specimens of the
weak and the strong REVs and (b) shows the position of the element, which is associated with the weak
REV (light gray). The cumulative deviatoric strain field corresponds to (c) 0.35% and (d) 8.0% axial strain.
Table VI. The relative time cost of the homoge-
neous CN-B simulation and IMP simulation.
Specimen Time cost [%]
Homogeneous CN-B specimen 100
IMP specimen 68
Comparing the IMP response with the homogeneous simulations shows that, after an identical
pre-peak behavior, the IMP yield to a smoother response at and beyond the peak. The weak REV has
not affected the strength capacity of IMP specimen, as the slight reduction in the strength peak can
be neglected comparing with the corresponding inhomogeneous simulations in the previous section.
Figure 12(c) and (d) displays the cumulative deviatoric strain field corresponding to 0.35% and
8% axial strain, respectively. As anticipated, it can be observed that the element associated with the
weak REV (CN-A) has triggered strain localization at 0.35% axial strain. Moreover, the consequent
shear band passes through the imperfect element as can be seen in Figure 12(d).
It is worthy to mention that the introduction of material imperfection into simulation specimen
has a positive effect on the computation time. Table VI displays the time cost associated to both
the IMP and the homogeneous CN-B simulations. The introduction of material imperfection has
reduced time cost about 35%. This matter can be attributed to the capability of the imperfect element
to trigger strain localization and to define the shear band locus along the deformation process. This
issue has been extensively studied in [35]. These observations give a first hint to the capability of
using different DEM REVs with the aforementioned specifications to improve the stability of the
FEM DEM simulations.
8. CONCLUSION
The consequences of using several DEM REVs for the microscale simulation in the framework of
multiscale FEM DEM modeling approach have been investigated in this work. The DEM assem-
blies were generated having the same material properties but with geometrical alterations. These
geometrical alterations have a strong effect on the mechanical behavior of the granular assembly as
shown by the pure DEM simulations. The performed homogeneous FEM DEM simulations have
shown a similar effect of geometry alteration, but with a much smoother response.
The performed inhomogeneous FEM DEM simulations reveal a clear evidence of the capability
of the weak DEM REV to trigger strain localization and to help the specimen at defining its shear
band pattern.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
Using a weak DEM REV to introduce local imperfection into homogeneous FEM DEM spec-
imen has shown a satisfactory effectiveness in terms of triggering strain localization; indeed, the
consequent shear band pattern passes through the imperfect element. The obtained results have given
a first hint to the ability of this technique to improve the stability of the problem and then to reduce
time cost.
In future work, it is worth to investigate the consequences of using a larger number of DEM
REVs in the FEM DEM simulations. A FE mesh of 50 elements has been used for the macroscale
level, and it is of significant importance to reevaluate the findings of this work with a finer FE mesh.
Also, it is recommended to repeat this study, adopting a regularization technique such as the second-
gradient theory, in order to mitigate the well-known effect of mesh dependency; this issue was out
of the scope of this study.
APPENDIX A: SENSITIVITY ANALYSIS
In this part, a sensitivity analysis is proposed to define a set of four parameters among others used in
the numerical FEMxDEM analysis. The four parameters are the inertial number (IN) and the level
of equilibrium (LE) conditions that are relevant to DEM scheme and the perturbation magnitude
(PM) and the precision level (EFand ED) that are relevant to FEM scheme. These four parameters
were selected assuming that they have an influential role on the simulation results and time cost.
The numerical experiments were carried out on two different FEMDEM simulations: the first was
homogeneously made of CN-A as an REV and the second was the RS-A.
A.1. DEM parameters: the inertial number and the level of equilibrium conditions
The IN is a physical parameter related to DE simulations that describes the strain-rate applied
on the granular assembly. As proposed in [10, 11], the quasi-static conditions are satisfied if the
IN 61.0E-3. Reducing the IN magnitude improves the quasi-static conditions; however, this reduc-
tion significantly increases time cost. Therefore, two different values of the IN have been examined:
IN D1.0E-3 and 1.0E-4. The LE parameter is the second DEM parameter that affects the quality
of the solution. The requirement of a higher LE conditions improves both the stability of the prob-
lem and the quality of the solution; however, this leads to a significant increase in the computational
time. The LE parameter was examined through two different levels, LE D2and 4 considering
the six levels of equilibrium conditions implemented in DEM code (Table I). The values of these
two parameters have been combined in four different combinations that are used in a set of biaxial
test simulations up to 2% axial strain. Table A.1 shows the four combinations and the associated
time cost.
Figure A.1(a) and (b) shows the global response of the two specimens corresponding to each
combination. Simulation number 1 combines the parameters with the stronger values (i.e., these
leading to higher quality of DEM modeling). Therefore, this simulation has been considered as a
Table A.1. The different combinations of parameters defined for the
inertial number and the level of equilibrium conditions sensitivity study
and the corresponding relative time cost, for 2%axialstrain.
Time cost [%]
No. IN LE PM ED EF
CN-A RS-A
11.0E-4 4 100 67
2 2 81 61
15.0E-6 0.01 0.01
31.0E-3 4 88 100
4 2 62 52
IN, inertial number; LE, level of equilibrium; PM, perturbation magnitude;
ED, displacement-based convergence criterion; EF, force-based conver-
gence criterion.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Figure A.1. A comparison of the global responses of the two specimens (a) CN-A and (b) RS-A
corresponding to four different combinations of the IN and the LE conditions, for 2% axial strain.
reference case. All cases have shown quite comparable pre-peak responses. However, simulations
numbers 3 and 4 show a softer response in the post-peak domain. These two simulations share
the same value of the IN, IN D1.0E-3, which refers to an important effect of this parameter on
the mechanical behavior of the FEM DEM specimen, specifically in the post-peak domain. In
contrast, simulation number 2 has demonstrated a good agreement in the behavior and a reduction
in time cost about 10% comparing to reference case. Therefore, the values of the IN and the LE in
combination number 2 have been adopted hereafter.
A.2. FEM parameters: the perturbation magnitude and the precision level
A.2.1. Perturbation magnitude. The derivation of the CTO relies on the computation of the REV
response on perturbed paths close to the actual incremental strain paths. The effects of the
perturbation magnitude on analysis results will be explored hereafter.
Loading is applied on the specimen as a vertical displacement of the top horizontal bound-
ary nodes. At the macro scale level (FEM), loading process is driven using an adaptive scheme,
which means that the displacement increment applied at each loading step (hereafter called ‘step
size’) might change during the computation. Therefore, a lower bound has to be specified for the
step size in order to avoid unlimited step size reduction in case of non convergence. This bound
was given the value 5.0E-5, which is equivalent to 2.5E-5 vertical strain. The determination of the
strain perturbation magnitude has to consider that the PM in the numerical derivation of the CTO
has to be defined as a fraction of the local strain rate increment, which depends on the adaptive
step size. Consequently, the lower bond of step size has been considered as a reference value, and
three values of PM have been examined: 5.0E-6, 10.0E-6, and 15.0E-6, which are fractions of the
lower bound. These values were used in the three combinations shown in Table A.2, in a set of
biaxial test simulations up to 0.4% axial strain (found to require 15 steps) using both specimens
CN-A and RS-A.
Figure A.2(a) and (b) displays the global responses of CN-A and RS-A specimens, respectively.
The three cases have led to comparable results with negligible changes in time cost. Such negligible
Table A.2. The three combinations defined for perturbation
magnitude sensitivity study.
No. IN LE PM ED EF
5 5.0E-6
6 1.0E-4 2 10.0E-6 0.01 0.01
7 15.0E-6
IN, inertial number; LE, level of equilibrium; PM, perturbation
magnitude; ED, displacement-based convergence criterion;
EF, force-based convergence criterion.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
Figure A.2. A comparison of the global responses of the two specimens (a) CN-A and (b) RS-A
corresponding to three different PM: 5.0E-6, 10.0E-6, and 15.0E-6, for 0.4% axial strain.
Table A.3. The different combinations of technical parameters defined
for precision level .EFand ED/sensitivity study and the corresponding
relative time cost, for 8% axial strain.
Time cost [%]
No. IN LE PM ED EF
CN-A RS-A
7 0.01 0.02 100 100
1.0E-4 2 15.0E-6
8 0.01 NA 25 19
IN, inertial number; LE, level of equilibrium; PM, perturbation magnitude;
ED, displacement-based convergence criterion; EF, force-based convergence
criterion.
changes in time cost can be attributed to the fact that perturbation method is applied in the first
steps when material behavior is quasi-elastic. Consequently, this parameter has a negligible influ-
ence on the solution and thus a PM equal to 15.0E-6 has been arbitrarily chosen to be used in
this work.
A.2.2. The precision level. The quality of the solution at the FEM scale is controlled by the preci-
sion parameters (convergence criteria) of the integration scheme. The required quality of the solution
is specified by two parameters: a force-based parameter, EF, representing the ratio of the applied
and the reaction forces norm, together with a displacement-based parameter, ED, accounting for the
decrement of the displacement field norm with the evolution of the iterative process [28].
Guo and Zhao [20] reported the use of displacement-based criterion solely as a convergence
indicator with ED60:01 for a similar multiscale FEM DEM simulation. This assumption was
based on a common consideration in the commercial FEM software such as GEO5, ANSYS,and
ABAQUS. In this study, two different precision combinations were defined: EDjEFD0:01j0:02 and
0.01jNA (i.e., double convergence criteria with EDD0:01 and EFD0:02, and single convergence
criterion with EDD0:01 disregarding the EFcriterion). In the sequel, the former is referred as
higher precision combination (HPC) and the latter lower precision combination (LPC). These two
combinations were used in several biaxial test simulations up to 8% axial strain on both CN-A and
RS-A specimens. Table A.3 depicts the performed simulations and the associated time cost.
Figure A.3 shows the global responses of both simulations, CN-A and RS-A, with HPC and
LPC. Generally speaking, no important change can be observed in the global responses of the spec-
imens when the force-based criterion is considered together with the displacement-based criterion.
However, neglecting the force-based criterion induces a significant reduction, about 75–80%, in
time cost.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

G. SHAHIN ET AL.
Figure A.3. A comparison of the global responses of two specimens (a) CN-A and (b) RS-A at two different
precision combinations: EDjEFD0:01jNA and 0.01j0.02, for 8% axial strain; ED, displacement-based
convergence criterion; EF, force-based convergence criterion.
Table A.4. The adopted computational parameters
values based on the performed sensitivity study.
IN LE PM EDjEF
1.0E-4 2 15.0E-6 0.01jNA
IN, inertial number; LE, level of equilibrium; PM,
perturbation magnitude; ED, displacement-based
convergence criterion; EF, force-based convergence
criterion.
As a result of this sensitivity analysis, the four parameters that are used in the current work are
defined and summarized in Table A.4.
ACKNOWLEDGEMENTS
Ghassan Shahin acknowledges the financial support of the European Education, Audiovisual and Culture
Executive Agency (EACEA), under the Erasmus Mundus Master in Earthquake Engineering and Engineer-
ing Seismology (MEEES) program. Also, he is pleased to acknowledge with thanks the helpful discussions
with Abraham Van den Eijnden. Laboratory 3SR is a part of the LabEx Tec 21 (Investissements d’Avenir –
grant agreement no. ANR-11-LABX-0030).
REFERENCES
1. Kaneko K, Terada K, Tyoya T, Kishino Y. Global-local analysis of granular media in quasi-static equilibrium.
International Journal of Solids and Structures 2003; 40(15):4043–4069.
2. Meier H, Steinmann P, Kuhl E. On the multiscale computation of confined granular media. In ECCOMAS Multidis-
ciplinary Jubilee Symposium, Computational Methods in Applied Sciences, Vol. 14, Eberhardsteiner J, Hellmich C,
Mang H, Priaux J (eds). Springer: Netherlands, 2009; 121–133.
3. Miehe C, Dettmar J, Zah D. Homogenization and two-scale simulations of granular materials for different
microstructural constraints. International Journal for Numerical Methods in Engineering 2010; 83(8–9):1206–1236.
4. Nitka M, Combe G, Dascalu C, Desrues J. Two-scale modeling of granular materials: a DEM-FEM approach.
Granular Matter 2011; 13(3):277–281.
5. Andrade J, Avial C, Hall S, Lenoir N, Viggiani G. Multiscale modelling and characterization of granular matter: from
grain kinematics to continuum mechanics. Journal of the Mechanics and Physics of Solids 2011; 59(2):237–250.
6. Guo N, Zhao J. A hierarchical model for cross-scale simulation of granular media. AIP Conference Proceedings
2013; 1542(1):1222–1225.
7. Wellmann C, Lillie C, Wriggers P. Homogenization of granular material modeled by a three-dimensional discrete
element method. Computers and Geotechnics 2008; 35(3):394–405.
8. Geers M, Kouznetsova V, Brekelmans W. Multi-scale computational homogenization: trends and challenges. Journal
of Computational and Applied Mathematics 2010; 234(7):2175–2182. Fourth International Conference on Advanced
COmputational Methods in ENgineering (ACOMEN 2008).
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme

INFLUENCE OF REV VARIABILITY IN DOUBLE-SCALE FEM DEM ANALYSIS
9. Cundall P, Ruest M, Guest A, Chitombo G. Evaluation of shemes to improve the efficiency of a complete model of
blasting and rock fractures. In Numerical Modeling in Micromechanics via Particles Methods, (Proceedings of the
1st International PFC Symposium), Konietzky H (ed.). Balkema: Gelsenkirchen, Germany, 2003; 107–115.
10. Radjai F, Dubois F. Discrete-element Modeling of Granular Materials. Wiley, 2011.
11. Roux J, Combe G. How granular materials deform in quasistatic conditions. In IUTAM-ISIMM Symposium on
Mathematical Modeling and Physical Instances of Granular Flows, AIP Conference Proceedings, Vol. 1227. Reggio
Calabria: Italy, 2010; 260–270.
12. Cundall P. A discontinuous future for numerical modelling in geomechanics? Proceedings of the ICE – Geotechnical
Engineering 2001; 149(1):41–47.
13. Terada K, Kurumatani M. Two-scale diffusion-deformation coupling model for material deterioration involoving
micro-crack propagation. International Journal for Numerical Methods in Engineering 2010; 83(4):426–451.
14. Miehe C, Dettmar J. A framework for micro–macro transitions in periodic particle aggregates of granular materials.
Computer Methods in Applied Mechanics and Engineering 2004; 193(35):225–256.
15. Meier H, Steinmann P, Kuhl E. Towards multiscale computation of confined granular media. Technische Mechanik
2008; 28(1):32–42.
16. Avial C, Andrade J. Advances in multiscale modelling and characterization of granular matter. Procedia IUTAM
2012; 3(0):157–171.
17. Nitka M, Bilbie B, Combe G, Dascalu C, Desrues J. A micro–macro (DEM–FEM) model of the behavior of granular
solids. In 1st International Symposium on Computational Geomechanics (ComGeo I), Juan-les-Pins, France, 2009;
38–48.
18. Nguyen TK, Combe G, Caillerie D, Desrues J. Modeling of a cohesive granular materials by a multi-scale approach.
AIP Conference Proceedings, Vol. 1542, Sydney, Australia, 2013; 1194–1197.
19. Nguyen T, Combe G, Caillerie D, Desrues J. FEM DEM modelling of cohesive granular materials: numerical
homogenisation and multi-scale simulations. Acta Geophysica 2014; 62(5):1109–1126.
20. Guo N, Zhao J. A coupled FEM/DEM approach for hierarchical multiscale modelling of granular media. Interna-
tional Journal for Numerical Methods in Engineering 2014; 99(11):789–818.
21. Desrues J, Nguyen T, Combe G, Caillerie D. FEM DEM multi-scale analysis of boundary value problems involving
strain localization. In Bifurcation and Degradation of Geomaterials in the New Millennium, Chau KT, Zhao J (eds).,
Springer Series in Geomechanics and Geoengineering. Springer International Publishing, 2015; 259–265.
22. Zhao J, Guo N. The interplay between anisotropy and strain localisation in granular soils: a multiscale insight.
Géotechnique 2015; 65(8):642–656.
23. Guo N, Zhao J. Multiscale insights into classical geomechanics problems. International Journal for Numerical and
Analytical Methods in Geomechanics 2015. DOI: 10.1002/nag.2406, In press.
24. Andò E. Étude experimentale de l’évolution de la microstructure d’un milieu granulaire sous chargement mécanique
a l’aide de la tomographie rayons x. PhD Thesis, Docteur de L’Université de Grenoble, 2013.
25. Weber J. Recherches concernant les contraintes intergranulaires dans les milieux pulverulents, Bull. Liaison des Ponts
et Chaussees 1966; 22:1–20.
26. Coenen E, Kouznetsova V, Bosco E, Geers M. A multi-scale approach to bridge microscale damage and macroscale
failure: a nested computational homogenization–localization framework. International Journal of Fracture 2012;
178(1–2):157–178.
27. Bosco E, Kouznetsova VG, Geers MGD. Multi-scale computational homogenization–localization for propagating
discontinuities using X-FEM. International Journal for Numerical Methods in Engineering 2015; 102(3–4):496–527.
28. Charlier R. Approache unifiée, de quelques problémes non linéaires de mécanique des milieux continus par la
méthode des éléments finis. PhD Thesis, University of Liège, France, 1987.
29. Zienkiewicz O, Taylor R, Zhu J. The Finite Element Method: Its Basis and Fundamentals. Butterworth-Heinemann,
2005.
30. Ypma T. Historical development of the Newton–Raphson method. SIAM Review 1995; 37(4):531–551.
31. Allen MP, Tildesley DJ. Computer Simulation of Liquids (Reprint edn.) Oxford Science Publications, Oxford
University Press: USA, 1989.
32. Desrues J, Viggiani G. Strain localization in sand: an overview of the experimental results obtained in Grenoble
using stereophotogrammetry. International Journal for Numerical and Analytical Methods in Geomechanics 2004;
28(4):279–321.
33. Vardoulakis I. Shear band inclination and shear modulus of sand in biaxial tests. International Journal of Numerical
and Analytical Methods in Geomechanics 1980; 4(2):103–119.
34. Saint-Cyr B, Szarf K, Voivret C, Azéma E, Richefeu V, Delenne JY, Combe G, Nouguier-Lehon C, Villard P,
Sornay P, Chaze M, Radjai F. Particle shape dependence in 2D granular media. Europhysics Letters - EPL 2012;
98(4):44008.
35. Besuelle P, Chambon R, Collin F. Switching deformation modes in post-localization solutions with a quasibrittle
material. Journal of Mechanics of Materials and Structures 2006; 3:1115–1134.
Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng (2016)
DOI: 10.1002/nme